An Exploration of Artificial Intelligence Assisted Strategies in English Reading Teaching

IRT, also known as Latent Trait Theory or Item Curve Theory, is a modern testing theory proposed to overcome the shortcomings of classical testing theory.

In the mathematical model of IRT, it is assumed that the learning ability of learner i can be measured, denoted by θ_i. Each item, i.e., learner exercise test, will have a difficulty coefficient of Exercise j denoted as β_j. The probability of a learner answering a question correctly by answering an exercise numbered j on a test is expressed using probability P(a), P(a) determined by the item response function, which is characterized by the following: if a learner has a higher learning ability, this learner has a higher probability of answering a question correctly: if the question answered is more difficult, the probability of the learner answering the question correctly is lower. In item response theory models, logistic regression models are commonly used as item response functions: (1) P(a)=sigmoid(θ−βj)=11+exp(−(θ−βj))

The item response functions for the two-parameter logistic regression IRT model are as follows: (2) Pi(θi)=11+exp(−1.7aj(θi−bj))

In the two-parameter logistic regression model, a_j denotes the discriminant parameter of item j, which represents the learner’s discrimination ability, and b_j denotes the discriminant parameter of item j, which represents the difficulty coefficient of the exercise.

IRT is designed for educational testing environments, and the model assumes that learners’ abilities do not change in the testing environment; however, in knowledge tracking, students’ knowledge status changes over time, and therefore, the initial IRT cannot be used directly for knowledge tracking tasks.

Knowledge tracking models considering forgetting behavior can be divided into two categories, knowledge tracking models based on the extension of the BKT model and deep knowledge tracking models [25].

In the proposed BKT model considering forgetting behavior, the default is that the learner forgets the previously learned knowledge after a few days of learning, and a new parameter “forgetn” is introduced to indicate the forgetting rate on a certain day after the learning interaction, but this model cannot simulate the forgetting behavior on a smaller scale.

On the basis of the BKT model considering forgetting behavior, a knowledge tracking model is proposed to estimate the probability of forgetting through the exercise sequence, for example, there is an example of exercises with exercise sequence A₁–A₂–B₁–A₃–B₂–B₃–A₄ denoting and knowledge point A,B, assuming that the probability of forgetting between A₁,A₂ is F, then the probability of forgetting between A₂ and A₃ is 1–(1–F)², and that the probability of forgetting between A₃ and A₄ is 1–(1–F)³. However, the model does not take into account the interval between the before and after answering sequences.

Considering the three forgetting behaviors, the DKT model is extended to propose a DKT model with forgetting ability, and the model’s structure for modeling forgetting behavior is shown in Figure 2.

Figure 2.

DKT-forgetful model structure

First, the model expresses the exercise interaction sequence x_t = (q_t,r_t) embedding as the exercise embedding vector v_t, and the three forgetting factors embedding as the vector c_t, and operates the two to obtain the embedding vector vtc after the forgetting process: (3) vtc=f(vt,ct)

Here f denotes one of the tandem, product or tandem product operations.

Then, the knowledge mastery state value h_t and the knowledge mastery state value after forgetting treatment are calculated htc . Finally, the probability of answering the question correctly next time is predicted based on htc : (4) ht=RNN(vtc,ht−1) (5) htc=f(ht,ct+1) (6) yt=sigmoid(Wouthtc+bout)

Deep-IRT’s model combines IRT and DKVMN model, which retains the performance of DKVMN model, provides mental explanations for both learners and exercises, and improves the interpretability of deep knowledge tracking model. In the DKVMN model, for an exercise q_t received at timestamp t, a feature vector f_t is generated after model training, which contains the learner’s knowledge mastery state and exercise embedding information for exercise q_t. The Deep-IRT model uses a neural network to build a student ability network layer and an exercise difficulty network layer, uses f_t vectors to infer the learner’s learning ability on exercise q_t, and uses the Exercise Embedding Vector q_t to compute the difficulty of the exercises: (7) θtj=tanh(Wθft+bθ) (8) βj=tanh(Wβqt+bβ)

Here θ_ti and β_j denote the learning ability and the difficulty of the exercise for knowledge point j at time stamp t, respectively, and then the probability that the learner will answer the question correctly is calculated p_t based on the learning ability and the difficulty of the exercise: (9) pt=sigmoid(3.0*θtj−βj)

The output of the learning ability network layer was scaled up by multiplying it by a factor of 3.0 for practical reasons, and the item response theory that was refined in this model can be applied to other models.

CF-DKD is a supervised learning method that tracks the learner’s knowledge proficiency through the learner’s sequence of test responses on each assessment.

The task description of CF-DKD is as follows: learner history interaction data is a sequence of learner responses to test questions at different moments, denoted by X = {x₁,x₂,…,x_t–1}. where the answer-response record, cell x_t = (e_t,r_t,f_t,l_t), is a quaternion indicating that a particular learner executes trial e_t,(e_t ∈ {e1,…,e|E|}) once per trial at a certain moment t. The incorrect and correct responses are r_t, which is a binary variable r_t = {0,1}. If the learner answers correctly, r_t = 1. Otherwise, r_t = 0. In addition, we use f_i,l_i to denote the cognitive factors forgetting and learning, respectively.

The solution is shown in Figure 3. We propose a dynamic knowledge diagnostic method (CF-DKD) incorporating cognitive features to track changes in learners’ potential knowledge mastery status.

Figure 3.

Solution

The complete one-step CF-DKD framework is shown in Fig. 4. CF-DKD is a time-series model that utilizes a memory network to dynamically store potential knowledge capabilities, and it provides a cognitive rule-based interface for reading and writing hidden information.

Figure 4.

Complete step CF-DKD framework

Based on the coding basis of extrinsic features, CF-DKD achieves the modeling of cognitive laws of learning and forgetting by classifying extrinsic knowledge features, interaction features, behavioral features, and temporal features according to cognitive laws, i.e., the modeling of cognitive laws of learning and forgetting. By introducing forgetting and learning factors to improve learners’ knowledge mastery [26]. The forgetting and learning features are shown in Figure 5.

Figure 5.

Cognitive feature extraction

We use the key-value matrix pairs M_cw and M_sw to store knowledge mastery states in CF-DKD, instead of using a single hidden layer to store knowledge mastery states in the more traditional DKT model. M_ek is an invariant key matrix that stores potential knowledge, and M_sw is a dynamic value matrix that stores each learner’s knowledge proficiency. These two memory matrices have the same slots, each representing a potential knowledge. The memory layer consists of two steps:

In order to obtain the relevance of the current test question to the potential knowledge, we use the attention mechanism by calculating the inner product vector k_i of the embedding vector of the current test question with the key matrix M_ece as follows: (11) wt(i)=softmax(kiTMex∫′(i))

where softmax(zi)=ezi∑i=1nez, and dw_i(i) ∈ [0,1].

First, given test questions and corresponding response tuples as inputs, the external memory approach uses wipe gates and add gates to update the hidden layer, ignoring the effect of long-term cognitive dependence on learners’ knowledge acquisition. Although various forgetting variables are added to the DKT+F model to affect the updating process, only one hidden layer cannot accurately represent the real hidden learning process of RNNs. Finally, we use an additional key-value mnemonic to store potential states. Thus, to balance the learning and forgetting factors, we propose two gates inspired by the thresholding mechanism to adaptively fuse the two features. Examples are the forgetting gate from LSTM, the updating gate from GRU, and the adding and deleting gate from GKT.

In our CF-DKD update method, the forgetting gate F_t controls the information to be deleted from the post-new-trial value matrix Mstut−1 , which contains the embedded current responses 3 v_t and the long-term forgetting factor f_t. The temporary forgetting vector f˜t is generated by combining the learner’s responses v_t(v_t ∈ ℝ^|d_v|) and forgetting features f_t(f_t ∈ ℝ^|d_f|). The forgetting information F_t can be computed by the fully connected layer with activation as: (12) f˜t=φ(ft,νt) (13) Fι=sigmoid(FT⋅f˜t+bf)

f_t = φ(Δs_t,Δr_t) and φ(·) are the integration functions that will be introduced in section “(VI) Prediction”. F^τ is a weight matrix F^T ∈ ℝ(d_r + d_f)p(d_r+d_f), where each element is a scalar between 0 and 1.

Similar to the forgetting gate, the learning gate L_t controls which knowledge needs to be enhanced in the current knowledge state Mstut−1 through the current response v_t and the long-term memory of the learning factor t_t. Therefore, we use the same integration method to integrate these two factors into a temporary learning vector l˜t . We can obtain the learning information through a fully connected layer with tanh activation L_t. The vector L_t is computed as (14) l˜t=φ(ΔCt,νt) (15) Lt=tanh(LT⋅l˜t+b1)

where is the L^T-weight matrix.

However, not all information has the same effect of weakening or strengthening prior information for the gate mechanism. We therefore multiply Mstut−1(i) with the correlation weight w_t to solve the problem of correlated trials during reading. Thus, the value memory matrix Mstut−1(i) is updated in the following way: (17) Mstut=Mstut−1(i)[ 1−wt(i)Ft ]+wt(i)Lt

How can we integrate various factors to predict learners’ performance on the next trial?

Based on the students’ scores in the reading section, the candidates’ reading scores were converted into levels according to the score-level conversion criteria. The CEDT test reading scores were converted into levels according to the score-level conversion criteria and the ratio of the number of students as shown in Table 1. Reading paper score (0,30) can be converted into 11 levels, 010-025 level students into the C level class (beginner), 030-040 level into the B level class (intermediate), 045-060 level into the A level class (advanced) for categorized teaching. The majority of candidates’ total scores are concentrated in the (13,15) and (16,18) bands, corresponding to the 035 level (22.00%) and the 040 level (17.8%).

The total reading scores of the overall students were analyzed with descriptive statistics using the SPSS statistical package and the descriptive statistics of the overall reading scores of the students are shown in Table 2. The sample size of the cohort was large (N=2668), the full distance of the sample distribution was wide (Range=29), and the inter-individual variation was large (Std=4.75), which is encompassing and representative, and the reading ability of the subject sample as a whole was good (Mean=15.82, Mode=16).

A histogram of the distribution of total student reading scores is shown in Figure 6. The normal Q-Q probability plot of total reading scores is shown in Figure 7. Observing the graph, it can be seen that the overall distribution of students’ total reading scores slightly deviates from the normal distribution line, with a skewness index of -0.175 (absolute value less than 1), which is slightly negatively skewed.

Figure 6.

The students read the total distribution histogram

Figure 7.

Read the total positive state of the q probability diagram

In this section, a series of cognitive diagnostic analyses of the test’s empirical data are conducted using the G-DINA model. Multilevel diagnostics are conducted for students in general, for groups of different levels and for individual students, while the relationships between reading attributes are discussed to make a case for the applicability of the G-DINA model in language testing.

The overall attribute patterns and distribution ratio of students are shown in Figure 8. The results show that the two patterns, (11111) and (00000), are the most prevalent patterns of attribute mastery. This finding is consistent with other cognitive diagnostic studies. The phenomenon is determined by the high positive correlation between the attributes, which means that the mastery of one of the attributes tends to predict the mastery of the other, and vice versa.

Figure 8.

The overall property pattern and distribution ratio of students

The differences in the probability of mastery of the reading attributes for the three level groups are shown in Table 6. Compared with the level A group, the level B and C groups have some differences in the probability of mastery of each attribute, with the most significant differences in attributes A4 and A5. For both the B and C level groups, the attribute with the worst probability of mastery is A5, which is significantly lower than the average probability of mastery of the other attributes, indicating that remedial learning in this area should be strengthened in the primary and intermediate student groups.

In this section, 10 students were randomly selected from the cohort to observe their scoring patterns and attribute mastery. The attribute mastery of the 10 individual students is shown in Table 7 (“0” means that the question was answered incorrectly and “1” means that the question was answered correctly). As can be seen from the table, although the total scores of this group of students were the same, and their reading comprehensiveness was the same as grade B, their scoring patterns on reading varied greatly, and none of them was the same. At the same time, the mastery of the five attributes involved in reading was completely different. Clearly, there are qualitative differences in the knowledge and cognitive structures of students with the same total and composite scores. In traditional tests, these qualitative differences among individual students are often obscured by a total score.

Since Class A and Class B are two independent individuals when comparing their reading performance horizontally, the independent samples t-test should be used to determine whether there is a significant difference between the performance of the two classes in terms of statistical significance, and the results of the independent samples t-test for Class A and Class B are shown in Table 12. From the test results of the table, the observed value of 0.382 by F-test and the probability value of 0.529 (>0.05) indicate that there is no significant difference between Class A and Class B in terms of their performance in English reading achievement.

Figure 9.

The midterm reading strategy mastered the results of the questionnaire survey

The results of the questionnaire on the mastery of reading strategies in Class B at midterm are shown in Figure 10. From the figure, it can be seen that the three best-performing strategies of class B students are cultural knowledge background, topic sentence, and pre-reading metacognitive strategy, and the worst-performing ones are mind mapping, classification, selective marking, and visualization. It can be concluded that the students in Class A and Class B are similar in the variability of reading strategy mastery, and the worst and best reading strategies of the two classes are basically the same.

Figure 10.

The midterm reading strategy mastered the results of the questionnaire survey

In order to verify that the method of this paper can effectively improve students’ reading levels, the researcher conducted a questionnaire survey to assess students’ mastery of reading strategies and two dimensions of English reading scores.

The results of the questionnaire survey on the mastery of reading strategies using the method of this paper and the traditional method are shown in Figures 11 and 12. Comparison with the pre-experimental results can be visualized that the overall level of both groups improved after the experiment, but the experimental group improved more obviously, and most of the scores of the traditional group were below 4.0. The before and after of the experimental group showed significant differences, while the before and after of the traditional group yielded very similar results. This indicates that from the visual descriptive statistics there is a significant difference between the experimental group and the control group after the experiment, and that the experimental group’s level of mastery of reading strategies increased more than that of the control group.

Figure 11.

The reading strategy mastered the results of the questionnaire survey

Figure 12.

The reading strategy mastered the results of the questionnaire survey

In order to more scientifically verify the effect of this paper’s method on reading achievement, this study used cross-comparison analysis to compare the experimental group’s midterm and final reading scores, as well as the experimental group’s final reading scores with those of the control group. The comparison of the experimental group’s midterm and final reading scores can prove whether the experimental group learning with this paper’s method effectively improves the reading scores, while the comparison of the experimental group’s final reading scores with the control group can filter out the influence of other variables besides this paper’s method, thus proving that this paper’s method can effectively improve students’ reading scores. The descriptive statistical trilinear table of the final reading achievement of the experiment using the method of this paper and the traditional teaching method is shown in Table 13 and Table 14. The table shows that the mean scores of the final reading scores of the experimental and control classes were 74.88 and 69.25 with standard deviations of 12.322 and 19.527 respectively. A comparison of the size of the numbers leads to the conclusion that the experimental class taught using the methodology of this paper had higher final reading scores than the control class that was taught according to the traditional model, and that there was even less disparity within the class.